Skip to content

Latest commit

 

History

History
104 lines (89 loc) · 3.63 KB

README.md

File metadata and controls

104 lines (89 loc) · 3.63 KB
Raw

Aligning Audiobooks

Audiobooks are large speech files that usually contain a lot of "sentences". Speech recognition and synthesis training systems require well aligned, relatively short examples of audio and text [1]. To segment them into smaller chunks. Follow the instructions below to align your text:

Setup and installation

  1. Clone and install CMU_Wilderness. You should stop at the make dependencies section.
  2. Set the environment variables in [cmu_wilderness_directory]/festvox_env_settings
  3. cd path/to/cmu_wilderness
  4. Make a directory called txt inside cmu_wilderness and copy your transcriptions files in txt format there.
  5. Create directory called download inside cmu_wilderness and inside it create a file called flist. In flist list all the file names (without extensions) separated by a new line eg if you were aligning Mathew you will have :
MAT_001
MAT_025
MAT_026

The names of your audio and transcription files should be matching.

  1. Copy align.sh to your cmu_wilderness directory.

Preprocessing

Text normalization

  1. You might want to normalize the transcriptions first so that you have good alignments. Some of the things to look for are:
    • numbers
    • digits
    • money
    • dates and times
    • Acronyms and abbreviations
    • loan words

The Bible has a lot of numbers which might be written in digits. We created number-dictionaries for the data from Open.Bible. If you make a number dictionary for your language,please consider contributing.

  1. If you have an idea of the utterance/sentence boundaries that you want/want to enforce, you should add an empty line to enforce that. Eg for the Bible, you can separate by verses or for other datasets by a period.

Before

Audiobooks are large speech files that usually contain a lot of "sentences". To segment them into smaller chunks for use in TTS and ASR applications, follow the instructions below.

After

Audiobooks are large speech files that usually contain a lot of "sentences".


To segment them into smaller chunks for use in TTS and ASR applications, follow the instructions below.
  1. Create txt.done.data files. Do
bin/do_found txt_to_ttd txt/*

Audio preprocessing

  1. Convert the audio to wav : If the audio is in mp3 format, use the following command:
for i in [audio_path]*.mp3; do fname2=`basename $i .mp3`; ffmpeg -i $i $fname2.wav; done
  1. Converting the .wav files to 16K 16bit with:
sox orig/MAT_010.wav -r 16k -b 16 wav/MAT_010.wav

or to convert many files:

for f in orig/*.wav; do sox $f -r 16k -b 16 ../wav/$f;done

After this step, your audio files should be in wav folder inside cmu_wilderness.

Alignment

  1. Check that your directory has the following required structure:
   cmu_wilderness/
    align.sh
    bin/do_found
    build
        ...
    download/
        flist
    wav/
        MAT_001.wav
        MAT_002.wav
        MAT_003.wav
        ...
    txt/
        MAT_001.txt
        MAT_002.txt
        MAT_003.txt
        ...
    data/
        MAT_001.data
        MAT_002.data
        MAT_003.data
        ... 
    ...
  1. Run nohup bash align.sh & . This will run the alignment in the background. Check nohup.out file for progress. It can take upto 6 days depending on your compute power to fully align a 27hr audiobook(s).
  2. When the alignment is done, you will have the aligned files in the aligned folder.

Post-processing

  1. You can order your alignments and pick the best n alignments. [TO DO]

References

[1]: Alan W Black "CMU Wilderness Multilingual Speech Dataset" ICASSP 2019, Brighton, UK.